Chinese AI companies: a guide to China's top AI labs
A working map of the organisations that matter in China's AI industry — the frontier labs publishing open weights, the platform incumbents with distribution, and the chip and robotics firms that decide how much compute any of it can use. Updated as the landscape moves.
Frontier model labs
The five names that set the pace for Chinese frontier training runs and account for most of the open weights coming out of the country.
DeepSeek
深度求索Open weightsHangzhou · Founded 2023Spun out of quant fund High-Flyer. Became the reference point for frontier-grade reasoning models released under permissive licences at a fraction of Western training budgets.
- Key models:
- DeepSeek-V3, DeepSeek-R1
Alibaba (Qwen / Tongyi)
通义千问Mostly openHangzhou · Founded 2023The most prolific open-weight publisher in China. Qwen derivatives dominate community fine-tune leaderboards and ship inside Alibaba Cloud's enterprise stack.
- Key models:
- Qwen3 family, Qwen-VL, Qwen-Coder
Moonshot AI
月之暗面MixedBeijing · Founded 2023Built its reputation on ultra-long-context assistants for Chinese knowledge workers, then moved into open agentic and reasoning releases.
- Key models:
- Kimi K-series, Kimi chat assistant
Zhipu AI / Z.ai
智谱Mostly openBeijing · Founded 2019Tsinghua University spin-off with the deepest research lineage of the new cohort, and the strongest state and enterprise procurement footprint.
- Key models:
- GLM-4 family, CogVLM, CogVideo
ByteDance Seed
豆包大模型MixedBeijing · Founded 2023Consumer distribution at a scale nobody else in China has. Doubao is the country's most-used assistant; Seed handles the frontier research and media generation work.
- Key models:
- Doubao, Seed / Seedream, Seedance
Challengers and platform incumbents
Well-funded labs and big-tech divisions shipping competitive models, often specialised by modality or vertical.
MiniMax
稀宇科技MixedShanghai · Founded 2021Strong on long-context architectures and consumer media apps; Hailuo made it a serious name in video generation.
- Key models:
- MiniMax-M / Text-01, Hailuo video
StepFun
阶跃星辰Mostly openShanghai · Founded 2023Multimodal-first lab with a habit of open-sourcing audio and video models that few labs anywhere release.
- Key models:
- Step series, Step-Audio, Step-Video
01.AI
零一万物Mostly openBeijing · Founded 2023Founded by Kai-Fu Lee. Yi models put Chinese open weights on Western leaderboards early; the company has since tilted toward enterprise applications.
- Key models:
- Yi series
Baidu
文心一言MixedBeijing · Founded 2019First mover among the big platforms, with Kunlun AI chips and Apollo autonomous driving giving it a full-stack story.
- Key models:
- ERNIE 4.x / ERNIE-VL
Tencent Hunyuan
混元MixedShenzhen · Founded 2023Quietly one of the biggest open-weight contributors in 3D and video, with WeChat as a distribution channel nobody can match.
- Key models:
- Hunyuan LLM, Hunyuan3D, HunyuanVideo
Baichuan AI
百川智能Mostly openBeijing · Founded 2023Founded by Sogou's Wang Xiaochuan; has narrowed its focus toward medical and healthcare applications of foundation models.
- Key models:
- Baichuan series
iFlytek
科大讯飞ClosedHefei · Founded 1999The incumbent in Chinese speech and education tech, training on domestic Huawei Ascend silicon rather than Nvidia.
- Key models:
- Spark (Xinghuo)
SenseTime
商汤ClosedShanghai / Hong Kong · Founded 2014Computer-vision era giant that repositioned around foundation models and its own large compute clusters.
- Key models:
- SenseNova
Compute, chips and embodied AI
The hardware layer that decides how fast the model layer can move under export controls.
Huawei (Ascend / Pangu)
ClosedShenzhen · Founded 1987The centre of China's domestic compute answer to export controls: chips, CANN software stack, and its own enterprise model family.
- Key models:
- Ascend 910 accelerators, Pangu models
Cambricon
寒武纪ClosedBeijing · Founded 2016Listed pure-play AI chip designer, the main domestic alternative to Huawei silicon.
- Key models:
- Siyuan / MLU accelerators
Moore Threads
摩尔线程ClosedBeijing · Founded 2020Building general-purpose GPUs with a CUDA-adjacent software story aimed at domestic training clusters.
- Key models:
- MTT GPU series
Unitree
宇树科技MixedHangzhou · Founded 2016The most visible name in Chinese embodied AI, pairing cheap humanoid hardware with learned locomotion policies.
- Key models:
- G1 / H1 humanoids, Go quadrupeds
How to read this landscape
Three forces explain almost every move on this list. The first is compute: export controls push every lab toward efficiency work and toward domestic silicon, which is why Chinese labs so often lead on training-cost-per-capability rather than raw scale.
The second is openness as strategy. Publishing weights buys global mindshare that closed Chinese APIs cannot, so open releases have become the default distribution channel abroad — and the reason Qwen and DeepSeek derivatives are everywhere.
The third is distribution at home. ByteDance, Tencent, Alibaba and Baidu can put a model in front of hundreds of millions of users on day one; the newer labs have to win on model quality or on a specific vertical.
Frequently asked questions
Which Chinese AI companies release open-weight models?
DeepSeek, Alibaba's Qwen team, Zhipu AI, StepFun, 01.AI and Baichuan publish open weights regularly. Tencent Hunyuan and MiniMax open-source selected multimodal models while keeping their strongest systems proprietary.
Who are China's top open model labs?
By release cadence and downstream adoption: DeepSeek, Alibaba Qwen, Zhipu AI (GLM), Moonshot AI (Kimi) and StepFun.
How do Chinese AI companies get compute under export controls?
Pre-control Nvidia stockpiles, export-compliant variants, offshore cloud capacity, and increasingly domestic accelerators from Huawei Ascend, Cambricon and Moore Threads.